To combine or not to combine?

fancyfantasicΤεχνίτη Νοημοσύνη και Ρομποτική

7 Νοε 2013 (πριν από 3 χρόνια και 7 μήνες)

99 εμφανίσεις

To combine or not to combine?

Multiple classifier and prediction systems.

Prof. Bogdan Gabrys,

Computational Intelligence Research Group,

Bournemouth University


tomatic classification or prediction model building directly from data with a minimal or no human
supervision is already absolutely crucial in order to stay competitive and maximally exploit the data

quickly changing business environments. However, the
methodology for ensuring that created models (i.e.
classifiers, predictors, etc.) are as good as possible should be in place before using them with confidence.

No human supervision in model building also implies that one should use powerful enough techniqu
which can learn the data to any degree of accuracy. There are currently a lot of methods from soft
computing, machine learning, and statistics domains which, in principal, satisfy this requirement.

In the pattern recognition domain the examples include
the nearest
neighbour classifiers
, decision trees,
neural networks with sufficient number of hidden nodes, fuzzy if

then rules systems which are built
directly from data,
fuzzy techniques

including those

based on hyperbox fuzzy sets, Bayesian

networks or logical rule bases. From the statistical point of view most of these methods could be classified
as non
parametric models. The main challenge in such cases is to design a model building strategy which
would guard against overfitting of the tra
ining data or, in other words, would lead to a good generalisation

Individual classification/prediction models have recently been challenged by combined pattern recognition
or prediction systems, which often show better performance.


k will be concerned with pattern classification

and prediction

problems and in particular the
motivations and challenges behind using and designing multiple classifier systems. Starting with an attempt
to answer the question of why would one like to combin
e classifiers we will move to an overview of how to
combine them. This will be followed by discussing the issues of majority voting limits and an illustration of
the potential enhancement for theoretical error limits when using hierarchical multistage orga
nisations for
majority vote based combination systems. The talk will also cover the issues of classifier diversity,
classifier selection

various practical evolutionary algorithm based implementations of multistage,
multilevel and multidimensional selectio
fusion models

lternatives to

combining decisions of the predictors

in the context of neuro
fuzzy classifiers
together with
a number of algorithm independent learning approaches
attempting to answer the main question: To
combine or not to combine?
l also be

The results of extensive testing for many different benchmark problems will be shown throughout the